コメントを投稿するにはログインが必要です
ログインページへ記事ができると、章ごとの要約と、聴きどころへ飛べる目次がここに並びます。いまは音声で聴けます。
#2-70E. Achieving Expert-Level Diagnosis Extraction from EHRs using LLM Prompting (Ann Rheum Dis, 2025)
2025年12月13日 05:00·4分36秒
How can we efficiently extract high-fidelity diagnostic information from vast electronic health records (EHRs) without the intensive manual annotation required by traditional supervised machine learning models? This episode discusses "Work smarter, not harder," a study that leverages the zero-shot inference capabilities of Large Language Models (LLMs)—specifically a locally hosted Deep-Seek R1 model—to overcome the data annotation bottleneck while adhering to GDPR compliance. The researchers optimized the prompt design using rheumatologist instructions and Chain-of-Thought (CoT) reasoning to precisely identify diagnoses like Rheumatoid Arthritis (RA) and Osteoarthritis (OA) in clinical notes. The key finding is that, through optimal prompting, the LLM achieved expert-level accuracy (up to 99%) in extracting these diagnoses from EHRs, demonstrating generalizability across different centers and diseases. This capability significantly streamlines EHR research. However, it is crucial to recognize that inadequate prompting or LLM hallucinations can compromise patient safety, necessitating carefully designed prompts and robust fail-safes. For clinicians navigating the accelerated adoption of LLMs in medical practice, such as clinical summarization, this study provides an essential template for understanding proper use and limitations of these powerful tools, guiding safe and effective integration into daily practice.
Citation: Maarseveen TD, Selani D, Steinz N, ten Brinck R, Glas HK, Veris-van Dieren J, Reinders MJT, van den Akker EB, Knevel R. Work smarter, not harder: achieve expert-level diagnosis extraction from medical records with optimal prompting of large language models. Annals of the Rheumatic Diseases. 2025. DOI: 10.1016/j.ard.2025.11.001
Disclaimer: This audio summary is based on personal interpretation and does not guarantee the exact content of the original paper. Please refer to the original article for details.
